147e4937aSGao Xiang // SPDX-License-Identifier: GPL-2.0-only 247e4937aSGao Xiang /* 347e4937aSGao Xiang * Copyright (C) 2018 HUAWEI, Inc. 4592e7cd0SAlexander A. Klimov * https://www.huawei.com/ 547e4937aSGao Xiang */ 647e4937aSGao Xiang #include "zdata.h" 747e4937aSGao Xiang #include "compress.h" 847e4937aSGao Xiang #include <linux/prefetch.h> 947e4937aSGao Xiang 1047e4937aSGao Xiang #include <trace/events/erofs.h> 1147e4937aSGao Xiang 1247e4937aSGao Xiang /* 139f6cc76eSGao Xiang * since pclustersize is variable for big pcluster feature, introduce slab 149f6cc76eSGao Xiang * pools implementation for different pcluster sizes. 159f6cc76eSGao Xiang */ 169f6cc76eSGao Xiang struct z_erofs_pcluster_slab { 179f6cc76eSGao Xiang struct kmem_cache *slab; 189f6cc76eSGao Xiang unsigned int maxpages; 199f6cc76eSGao Xiang char name[48]; 209f6cc76eSGao Xiang }; 219f6cc76eSGao Xiang 229f6cc76eSGao Xiang #define _PCLP(n) { .maxpages = n } 239f6cc76eSGao Xiang 249f6cc76eSGao Xiang static struct z_erofs_pcluster_slab pcluster_pool[] __read_mostly = { 259f6cc76eSGao Xiang _PCLP(1), _PCLP(4), _PCLP(16), _PCLP(64), _PCLP(128), 269f6cc76eSGao Xiang _PCLP(Z_EROFS_PCLUSTER_MAX_PAGES) 279f6cc76eSGao Xiang }; 289f6cc76eSGao Xiang 299f6cc76eSGao Xiang static void z_erofs_destroy_pcluster_pool(void) 309f6cc76eSGao Xiang { 319f6cc76eSGao Xiang int i; 329f6cc76eSGao Xiang 339f6cc76eSGao Xiang for (i = 0; i < ARRAY_SIZE(pcluster_pool); ++i) { 349f6cc76eSGao Xiang if (!pcluster_pool[i].slab) 359f6cc76eSGao Xiang continue; 369f6cc76eSGao Xiang kmem_cache_destroy(pcluster_pool[i].slab); 379f6cc76eSGao Xiang pcluster_pool[i].slab = NULL; 389f6cc76eSGao Xiang } 399f6cc76eSGao Xiang } 409f6cc76eSGao Xiang 419f6cc76eSGao Xiang static int z_erofs_create_pcluster_pool(void) 429f6cc76eSGao Xiang { 439f6cc76eSGao Xiang struct z_erofs_pcluster_slab *pcs; 449f6cc76eSGao Xiang struct z_erofs_pcluster *a; 459f6cc76eSGao Xiang unsigned int size; 469f6cc76eSGao Xiang 479f6cc76eSGao Xiang for (pcs = pcluster_pool; 489f6cc76eSGao Xiang pcs < pcluster_pool + ARRAY_SIZE(pcluster_pool); ++pcs) { 499f6cc76eSGao Xiang size = struct_size(a, compressed_pages, pcs->maxpages); 509f6cc76eSGao Xiang 519f6cc76eSGao Xiang sprintf(pcs->name, "erofs_pcluster-%u", pcs->maxpages); 529f6cc76eSGao Xiang pcs->slab = kmem_cache_create(pcs->name, size, 0, 539f6cc76eSGao Xiang SLAB_RECLAIM_ACCOUNT, NULL); 549f6cc76eSGao Xiang if (pcs->slab) 559f6cc76eSGao Xiang continue; 569f6cc76eSGao Xiang 579f6cc76eSGao Xiang z_erofs_destroy_pcluster_pool(); 589f6cc76eSGao Xiang return -ENOMEM; 599f6cc76eSGao Xiang } 609f6cc76eSGao Xiang return 0; 619f6cc76eSGao Xiang } 629f6cc76eSGao Xiang 639f6cc76eSGao Xiang static struct z_erofs_pcluster *z_erofs_alloc_pcluster(unsigned int nrpages) 649f6cc76eSGao Xiang { 659f6cc76eSGao Xiang int i; 669f6cc76eSGao Xiang 679f6cc76eSGao Xiang for (i = 0; i < ARRAY_SIZE(pcluster_pool); ++i) { 689f6cc76eSGao Xiang struct z_erofs_pcluster_slab *pcs = pcluster_pool + i; 699f6cc76eSGao Xiang struct z_erofs_pcluster *pcl; 709f6cc76eSGao Xiang 719f6cc76eSGao Xiang if (nrpages > pcs->maxpages) 729f6cc76eSGao Xiang continue; 739f6cc76eSGao Xiang 749f6cc76eSGao Xiang pcl = kmem_cache_zalloc(pcs->slab, GFP_NOFS); 759f6cc76eSGao Xiang if (!pcl) 769f6cc76eSGao Xiang return ERR_PTR(-ENOMEM); 779f6cc76eSGao Xiang pcl->pclusterpages = nrpages; 789f6cc76eSGao Xiang return pcl; 799f6cc76eSGao Xiang } 809f6cc76eSGao Xiang return ERR_PTR(-EINVAL); 819f6cc76eSGao Xiang } 829f6cc76eSGao Xiang 839f6cc76eSGao Xiang static void z_erofs_free_pcluster(struct z_erofs_pcluster *pcl) 849f6cc76eSGao Xiang { 85cecf864dSYue Hu unsigned int pclusterpages = z_erofs_pclusterpages(pcl); 869f6cc76eSGao Xiang int i; 879f6cc76eSGao Xiang 889f6cc76eSGao Xiang for (i = 0; i < ARRAY_SIZE(pcluster_pool); ++i) { 899f6cc76eSGao Xiang struct z_erofs_pcluster_slab *pcs = pcluster_pool + i; 909f6cc76eSGao Xiang 91cecf864dSYue Hu if (pclusterpages > pcs->maxpages) 929f6cc76eSGao Xiang continue; 939f6cc76eSGao Xiang 949f6cc76eSGao Xiang kmem_cache_free(pcs->slab, pcl); 959f6cc76eSGao Xiang return; 969f6cc76eSGao Xiang } 979f6cc76eSGao Xiang DBG_BUGON(1); 989f6cc76eSGao Xiang } 999f6cc76eSGao Xiang 10047e4937aSGao Xiang /* how to allocate cached pages for a pcluster */ 10147e4937aSGao Xiang enum z_erofs_cache_alloctype { 10247e4937aSGao Xiang DONTALLOC, /* don't allocate any cached pages */ 1031825c8d7SGao Xiang /* 1041825c8d7SGao Xiang * try to use cached I/O if page allocation succeeds or fallback 1051825c8d7SGao Xiang * to in-place I/O instead to avoid any direct reclaim. 1061825c8d7SGao Xiang */ 1071825c8d7SGao Xiang TRYALLOC, 10847e4937aSGao Xiang }; 10947e4937aSGao Xiang 11047e4937aSGao Xiang /* 11147e4937aSGao Xiang * tagged pointer with 1-bit tag for all compressed pages 11247e4937aSGao Xiang * tag 0 - the page is just found with an extra page reference 11347e4937aSGao Xiang */ 11447e4937aSGao Xiang typedef tagptr1_t compressed_page_t; 11547e4937aSGao Xiang 11647e4937aSGao Xiang #define tag_compressed_page_justfound(page) \ 11747e4937aSGao Xiang tagptr_fold(compressed_page_t, page, 1) 11847e4937aSGao Xiang 11947e4937aSGao Xiang static struct workqueue_struct *z_erofs_workqueue __read_mostly; 12047e4937aSGao Xiang 12147e4937aSGao Xiang void z_erofs_exit_zip_subsystem(void) 12247e4937aSGao Xiang { 12347e4937aSGao Xiang destroy_workqueue(z_erofs_workqueue); 1249f6cc76eSGao Xiang z_erofs_destroy_pcluster_pool(); 12547e4937aSGao Xiang } 12647e4937aSGao Xiang 12799634bf3SGao Xiang static inline int z_erofs_init_workqueue(void) 12847e4937aSGao Xiang { 12947e4937aSGao Xiang const unsigned int onlinecpus = num_possible_cpus(); 13047e4937aSGao Xiang 13147e4937aSGao Xiang /* 13247e4937aSGao Xiang * no need to spawn too many threads, limiting threads could minimum 13347e4937aSGao Xiang * scheduling overhead, perhaps per-CPU threads should be better? 13447e4937aSGao Xiang */ 1350e62ea33SGao Xiang z_erofs_workqueue = alloc_workqueue("erofs_unzipd", 1360e62ea33SGao Xiang WQ_UNBOUND | WQ_HIGHPRI, 13747e4937aSGao Xiang onlinecpus + onlinecpus / 4); 13847e4937aSGao Xiang return z_erofs_workqueue ? 0 : -ENOMEM; 13947e4937aSGao Xiang } 14047e4937aSGao Xiang 14147e4937aSGao Xiang int __init z_erofs_init_zip_subsystem(void) 14247e4937aSGao Xiang { 1439f6cc76eSGao Xiang int err = z_erofs_create_pcluster_pool(); 14447e4937aSGao Xiang 1459f6cc76eSGao Xiang if (err) 1469f6cc76eSGao Xiang return err; 1479f6cc76eSGao Xiang err = z_erofs_init_workqueue(); 1489f6cc76eSGao Xiang if (err) 1499f6cc76eSGao Xiang z_erofs_destroy_pcluster_pool(); 1509f6cc76eSGao Xiang return err; 15147e4937aSGao Xiang } 15247e4937aSGao Xiang 15347e4937aSGao Xiang enum z_erofs_collectmode { 15447e4937aSGao Xiang COLLECT_SECONDARY, 15547e4937aSGao Xiang COLLECT_PRIMARY, 15647e4937aSGao Xiang /* 15747e4937aSGao Xiang * The current collection was the tail of an exist chain, in addition 15847e4937aSGao Xiang * that the previous processed chained collections are all decided to 15947e4937aSGao Xiang * be hooked up to it. 16047e4937aSGao Xiang * A new chain will be created for the remaining collections which are 16147e4937aSGao Xiang * not processed yet, therefore different from COLLECT_PRIMARY_FOLLOWED, 16247e4937aSGao Xiang * the next collection cannot reuse the whole page safely in 16347e4937aSGao Xiang * the following scenario: 16447e4937aSGao Xiang * ________________________________________________________________ 16547e4937aSGao Xiang * | tail (partial) page | head (partial) page | 16647e4937aSGao Xiang * | (belongs to the next cl) | (belongs to the current cl) | 16747e4937aSGao Xiang * |_______PRIMARY_FOLLOWED_______|________PRIMARY_HOOKED___________| 16847e4937aSGao Xiang */ 16947e4937aSGao Xiang COLLECT_PRIMARY_HOOKED, 1700b964600SGao Xiang /* 1710b964600SGao Xiang * a weak form of COLLECT_PRIMARY_FOLLOWED, the difference is that it 1720b964600SGao Xiang * could be dispatched into bypass queue later due to uptodated managed 1730b964600SGao Xiang * pages. All related online pages cannot be reused for inplace I/O (or 1740b964600SGao Xiang * pagevec) since it can be directly decoded without I/O submission. 1750b964600SGao Xiang */ 17647e4937aSGao Xiang COLLECT_PRIMARY_FOLLOWED_NOINPLACE, 17747e4937aSGao Xiang /* 17847e4937aSGao Xiang * The current collection has been linked with the owned chain, and 17947e4937aSGao Xiang * could also be linked with the remaining collections, which means 18047e4937aSGao Xiang * if the processing page is the tail page of the collection, thus 18147e4937aSGao Xiang * the current collection can safely use the whole page (since 18247e4937aSGao Xiang * the previous collection is under control) for in-place I/O, as 18347e4937aSGao Xiang * illustrated below: 18447e4937aSGao Xiang * ________________________________________________________________ 18547e4937aSGao Xiang * | tail (partial) page | head (partial) page | 18647e4937aSGao Xiang * | (of the current cl) | (of the previous collection) | 18747e4937aSGao Xiang * | PRIMARY_FOLLOWED or | | 18847e4937aSGao Xiang * |_____PRIMARY_HOOKED___|____________PRIMARY_FOLLOWED____________| 18947e4937aSGao Xiang * 19047e4937aSGao Xiang * [ (*) the above page can be used as inplace I/O. ] 19147e4937aSGao Xiang */ 19247e4937aSGao Xiang COLLECT_PRIMARY_FOLLOWED, 19347e4937aSGao Xiang }; 19447e4937aSGao Xiang 1955c6dcc57SGao Xiang struct z_erofs_decompress_frontend { 1965c6dcc57SGao Xiang struct inode *const inode; 1975c6dcc57SGao Xiang struct erofs_map_blocks map; 1985c6dcc57SGao Xiang 19947e4937aSGao Xiang struct z_erofs_pagevec_ctor vector; 20047e4937aSGao Xiang 20147e4937aSGao Xiang struct z_erofs_pcluster *pcl, *tailpcl; 20281382f5fSGao Xiang /* a pointer used to pick up inplace I/O pages */ 20381382f5fSGao Xiang struct page **icpage_ptr; 20447e4937aSGao Xiang z_erofs_next_pcluster_t owned_head; 20547e4937aSGao Xiang 20647e4937aSGao Xiang enum z_erofs_collectmode mode; 20747e4937aSGao Xiang 2086ea5aad3SGao Xiang bool readahead; 20947e4937aSGao Xiang /* used for applying cache strategy on the fly */ 21047e4937aSGao Xiang bool backmost; 21147e4937aSGao Xiang erofs_off_t headoffset; 21247e4937aSGao Xiang }; 21347e4937aSGao Xiang 21447e4937aSGao Xiang #define DECOMPRESS_FRONTEND_INIT(__i) { \ 2155c6dcc57SGao Xiang .inode = __i, .owned_head = Z_EROFS_PCLUSTER_TAIL, \ 2164398d3c3SWeizhao Ouyang .mode = COLLECT_PRIMARY_FOLLOWED, .backmost = true } 21747e4937aSGao Xiang 21847e4937aSGao Xiang static struct page *z_pagemap_global[Z_EROFS_VMAP_GLOBAL_PAGES]; 21947e4937aSGao Xiang static DEFINE_MUTEX(z_pagemap_global_lock); 22047e4937aSGao Xiang 2216f39d1e1SGao Xiang static void z_erofs_bind_cache(struct z_erofs_decompress_frontend *fe, 2221825c8d7SGao Xiang enum z_erofs_cache_alloctype type, 223eaa9172aSGao Xiang struct page **pagepool) 22447e4937aSGao Xiang { 2256f39d1e1SGao Xiang struct address_space *mc = MNGD_MAPPING(EROFS_I_SB(fe->inode)); 2265c6dcc57SGao Xiang struct z_erofs_pcluster *pcl = fe->pcl; 22747e4937aSGao Xiang bool standalone = true; 2286f39d1e1SGao Xiang /* 2296f39d1e1SGao Xiang * optimistic allocation without direct reclaim since inplace I/O 2306f39d1e1SGao Xiang * can be used if low memory otherwise. 2316f39d1e1SGao Xiang */ 2321825c8d7SGao Xiang gfp_t gfp = (mapping_gfp_mask(mc) & ~__GFP_DIRECT_RECLAIM) | 2331825c8d7SGao Xiang __GFP_NOMEMALLOC | __GFP_NORETRY | __GFP_NOWARN; 23481382f5fSGao Xiang struct page **pages; 23581382f5fSGao Xiang pgoff_t index; 23647e4937aSGao Xiang 2375c6dcc57SGao Xiang if (fe->mode < COLLECT_PRIMARY_FOLLOWED) 23847e4937aSGao Xiang return; 23947e4937aSGao Xiang 24081382f5fSGao Xiang pages = pcl->compressed_pages; 24181382f5fSGao Xiang index = pcl->obj.index; 24281382f5fSGao Xiang for (; index < pcl->obj.index + pcl->pclusterpages; ++index, ++pages) { 24347e4937aSGao Xiang struct page *page; 24447e4937aSGao Xiang compressed_page_t t; 2451825c8d7SGao Xiang struct page *newpage = NULL; 24647e4937aSGao Xiang 24747e4937aSGao Xiang /* the compressed page was loaded before */ 24847e4937aSGao Xiang if (READ_ONCE(*pages)) 24947e4937aSGao Xiang continue; 25047e4937aSGao Xiang 25147e4937aSGao Xiang page = find_get_page(mc, index); 25247e4937aSGao Xiang 25347e4937aSGao Xiang if (page) { 25447e4937aSGao Xiang t = tag_compressed_page_justfound(page); 2550b964600SGao Xiang } else { 2560b964600SGao Xiang /* I/O is needed, no possible to decompress directly */ 2570b964600SGao Xiang standalone = false; 2580b964600SGao Xiang switch (type) { 2590b964600SGao Xiang case TRYALLOC: 2601825c8d7SGao Xiang newpage = erofs_allocpage(pagepool, gfp); 2611825c8d7SGao Xiang if (!newpage) 26247e4937aSGao Xiang continue; 2630b964600SGao Xiang set_page_private(newpage, 2640b964600SGao Xiang Z_EROFS_PREALLOCATED_PAGE); 2650b964600SGao Xiang t = tag_compressed_page_justfound(newpage); 2660b964600SGao Xiang break; 2670b964600SGao Xiang default: /* DONTALLOC */ 2680b964600SGao Xiang continue; 2690b964600SGao Xiang } 27047e4937aSGao Xiang } 27147e4937aSGao Xiang 27247e4937aSGao Xiang if (!cmpxchg_relaxed(pages, NULL, tagptr_cast_ptr(t))) 27347e4937aSGao Xiang continue; 27447e4937aSGao Xiang 275eaa9172aSGao Xiang if (page) 27647e4937aSGao Xiang put_page(page); 277eaa9172aSGao Xiang else if (newpage) 278eaa9172aSGao Xiang erofs_pagepool_add(pagepool, newpage); 27947e4937aSGao Xiang } 28047e4937aSGao Xiang 2810b964600SGao Xiang /* 2820b964600SGao Xiang * don't do inplace I/O if all compressed pages are available in 2830b964600SGao Xiang * managed cache since it can be moved to the bypass queue instead. 2840b964600SGao Xiang */ 2850b964600SGao Xiang if (standalone) 2865c6dcc57SGao Xiang fe->mode = COLLECT_PRIMARY_FOLLOWED_NOINPLACE; 28747e4937aSGao Xiang } 28847e4937aSGao Xiang 28947e4937aSGao Xiang /* called by erofs_shrinker to get rid of all compressed_pages */ 29047e4937aSGao Xiang int erofs_try_to_free_all_cached_pages(struct erofs_sb_info *sbi, 29147e4937aSGao Xiang struct erofs_workgroup *grp) 29247e4937aSGao Xiang { 29347e4937aSGao Xiang struct z_erofs_pcluster *const pcl = 29447e4937aSGao Xiang container_of(grp, struct z_erofs_pcluster, obj); 29547e4937aSGao Xiang int i; 29647e4937aSGao Xiang 297cecf864dSYue Hu DBG_BUGON(z_erofs_is_inline_pcluster(pcl)); 29847e4937aSGao Xiang /* 29947e4937aSGao Xiang * refcount of workgroup is now freezed as 1, 30047e4937aSGao Xiang * therefore no need to worry about available decompression users. 30147e4937aSGao Xiang */ 3029f6cc76eSGao Xiang for (i = 0; i < pcl->pclusterpages; ++i) { 30347e4937aSGao Xiang struct page *page = pcl->compressed_pages[i]; 30447e4937aSGao Xiang 30547e4937aSGao Xiang if (!page) 30647e4937aSGao Xiang continue; 30747e4937aSGao Xiang 30847e4937aSGao Xiang /* block other users from reclaiming or migrating the page */ 30947e4937aSGao Xiang if (!trylock_page(page)) 31047e4937aSGao Xiang return -EBUSY; 31147e4937aSGao Xiang 312f4d4e5fcSYue Hu if (!erofs_page_is_managed(sbi, page)) 31347e4937aSGao Xiang continue; 31447e4937aSGao Xiang 31547e4937aSGao Xiang /* barrier is implied in the following 'unlock_page' */ 31647e4937aSGao Xiang WRITE_ONCE(pcl->compressed_pages[i], NULL); 3176aaa7b06SGao Xiang detach_page_private(page); 31847e4937aSGao Xiang unlock_page(page); 31947e4937aSGao Xiang } 32047e4937aSGao Xiang return 0; 32147e4937aSGao Xiang } 32247e4937aSGao Xiang 323d252ff3dSYue Hu int erofs_try_to_free_cached_page(struct page *page) 32447e4937aSGao Xiang { 32547e4937aSGao Xiang struct z_erofs_pcluster *const pcl = (void *)page_private(page); 32647e4937aSGao Xiang int ret = 0; /* 0 - busy */ 32747e4937aSGao Xiang 32847e4937aSGao Xiang if (erofs_workgroup_try_to_freeze(&pcl->obj, 1)) { 32947e4937aSGao Xiang unsigned int i; 33047e4937aSGao Xiang 331cecf864dSYue Hu DBG_BUGON(z_erofs_is_inline_pcluster(pcl)); 3329f6cc76eSGao Xiang for (i = 0; i < pcl->pclusterpages; ++i) { 33347e4937aSGao Xiang if (pcl->compressed_pages[i] == page) { 33447e4937aSGao Xiang WRITE_ONCE(pcl->compressed_pages[i], NULL); 33547e4937aSGao Xiang ret = 1; 33647e4937aSGao Xiang break; 33747e4937aSGao Xiang } 33847e4937aSGao Xiang } 33947e4937aSGao Xiang erofs_workgroup_unfreeze(&pcl->obj, 1); 34047e4937aSGao Xiang 3416aaa7b06SGao Xiang if (ret) 3426aaa7b06SGao Xiang detach_page_private(page); 34347e4937aSGao Xiang } 34447e4937aSGao Xiang return ret; 34547e4937aSGao Xiang } 34647e4937aSGao Xiang 34747e4937aSGao Xiang /* page_type must be Z_EROFS_PAGE_TYPE_EXCLUSIVE */ 3485c6dcc57SGao Xiang static bool z_erofs_try_inplace_io(struct z_erofs_decompress_frontend *fe, 34947e4937aSGao Xiang struct page *page) 35047e4937aSGao Xiang { 3515c6dcc57SGao Xiang struct z_erofs_pcluster *const pcl = fe->pcl; 35247e4937aSGao Xiang 3535c6dcc57SGao Xiang while (fe->icpage_ptr > pcl->compressed_pages) 3545c6dcc57SGao Xiang if (!cmpxchg(--fe->icpage_ptr, NULL, page)) 35547e4937aSGao Xiang return true; 35647e4937aSGao Xiang return false; 35747e4937aSGao Xiang } 35847e4937aSGao Xiang 35987ca34a7SGao Xiang /* callers must be with pcluster lock held */ 3605c6dcc57SGao Xiang static int z_erofs_attach_page(struct z_erofs_decompress_frontend *fe, 36186432a6dSGao Xiang struct page *page, enum z_erofs_page_type type, 36286432a6dSGao Xiang bool pvec_safereuse) 36347e4937aSGao Xiang { 36447e4937aSGao Xiang int ret; 36547e4937aSGao Xiang 36647e4937aSGao Xiang /* give priority for inplaceio */ 3675c6dcc57SGao Xiang if (fe->mode >= COLLECT_PRIMARY && 36847e4937aSGao Xiang type == Z_EROFS_PAGE_TYPE_EXCLUSIVE && 3695c6dcc57SGao Xiang z_erofs_try_inplace_io(fe, page)) 37047e4937aSGao Xiang return 0; 37147e4937aSGao Xiang 3725c6dcc57SGao Xiang ret = z_erofs_pagevec_enqueue(&fe->vector, page, type, 37386432a6dSGao Xiang pvec_safereuse); 37487ca34a7SGao Xiang fe->pcl->vcnt += (unsigned int)ret; 37547e4937aSGao Xiang return ret ? 0 : -EAGAIN; 37647e4937aSGao Xiang } 37747e4937aSGao Xiang 3785c6dcc57SGao Xiang static void z_erofs_try_to_claim_pcluster(struct z_erofs_decompress_frontend *f) 37947e4937aSGao Xiang { 3805c6dcc57SGao Xiang struct z_erofs_pcluster *pcl = f->pcl; 3815c6dcc57SGao Xiang z_erofs_next_pcluster_t *owned_head = &f->owned_head; 38247e4937aSGao Xiang 383473e15b0SGao Xiang /* type 1, nil pcluster (this pcluster doesn't belong to any chain.) */ 384473e15b0SGao Xiang if (cmpxchg(&pcl->next, Z_EROFS_PCLUSTER_NIL, 385473e15b0SGao Xiang *owned_head) == Z_EROFS_PCLUSTER_NIL) { 38647e4937aSGao Xiang *owned_head = &pcl->next; 387473e15b0SGao Xiang /* so we can attach this pcluster to our submission chain. */ 3885c6dcc57SGao Xiang f->mode = COLLECT_PRIMARY_FOLLOWED; 389473e15b0SGao Xiang return; 390473e15b0SGao Xiang } 391473e15b0SGao Xiang 39247e4937aSGao Xiang /* 393473e15b0SGao Xiang * type 2, link to the end of an existing open chain, be careful 394473e15b0SGao Xiang * that its submission is controlled by the original attached chain. 39547e4937aSGao Xiang */ 39647e4937aSGao Xiang if (cmpxchg(&pcl->next, Z_EROFS_PCLUSTER_TAIL, 397473e15b0SGao Xiang *owned_head) == Z_EROFS_PCLUSTER_TAIL) { 39847e4937aSGao Xiang *owned_head = Z_EROFS_PCLUSTER_TAIL; 3995c6dcc57SGao Xiang f->mode = COLLECT_PRIMARY_HOOKED; 4005c6dcc57SGao Xiang f->tailpcl = NULL; 401473e15b0SGao Xiang return; 40247e4937aSGao Xiang } 403473e15b0SGao Xiang /* type 3, it belongs to a chain, but it isn't the end of the chain */ 4045c6dcc57SGao Xiang f->mode = COLLECT_PRIMARY; 40547e4937aSGao Xiang } 40647e4937aSGao Xiang 40783a386c0SGao Xiang static int z_erofs_lookup_pcluster(struct z_erofs_decompress_frontend *fe) 40847e4937aSGao Xiang { 40983a386c0SGao Xiang struct erofs_map_blocks *map = &fe->map; 4105c6dcc57SGao Xiang struct z_erofs_pcluster *pcl = fe->pcl; 41147e4937aSGao Xiang unsigned int length; 41247e4937aSGao Xiang 41364094a04SGao Xiang /* to avoid unexpected loop formed by corrupted images */ 4145c6dcc57SGao Xiang if (fe->owned_head == &pcl->next || pcl == fe->tailpcl) { 41547e4937aSGao Xiang DBG_BUGON(1); 4169e579fc1SGao Xiang return -EFSCORRUPTED; 41747e4937aSGao Xiang } 41847e4937aSGao Xiang 41987ca34a7SGao Xiang if (pcl->pageofs_out != (map->m_la & ~PAGE_MASK)) { 42047e4937aSGao Xiang DBG_BUGON(1); 4219e579fc1SGao Xiang return -EFSCORRUPTED; 42247e4937aSGao Xiang } 42347e4937aSGao Xiang 42447e4937aSGao Xiang length = READ_ONCE(pcl->length); 42547e4937aSGao Xiang if (length & Z_EROFS_PCLUSTER_FULL_LENGTH) { 42647e4937aSGao Xiang if ((map->m_llen << Z_EROFS_PCLUSTER_LENGTH_BIT) > length) { 42747e4937aSGao Xiang DBG_BUGON(1); 4289e579fc1SGao Xiang return -EFSCORRUPTED; 42947e4937aSGao Xiang } 43047e4937aSGao Xiang } else { 43147e4937aSGao Xiang unsigned int llen = map->m_llen << Z_EROFS_PCLUSTER_LENGTH_BIT; 43247e4937aSGao Xiang 43347e4937aSGao Xiang if (map->m_flags & EROFS_MAP_FULL_MAPPED) 43447e4937aSGao Xiang llen |= Z_EROFS_PCLUSTER_FULL_LENGTH; 43547e4937aSGao Xiang 43647e4937aSGao Xiang while (llen > length && 43747e4937aSGao Xiang length != cmpxchg_relaxed(&pcl->length, length, llen)) { 43847e4937aSGao Xiang cpu_relax(); 43947e4937aSGao Xiang length = READ_ONCE(pcl->length); 44047e4937aSGao Xiang } 44147e4937aSGao Xiang } 44287ca34a7SGao Xiang mutex_lock(&pcl->lock); 44347e4937aSGao Xiang /* used to check tail merging loop due to corrupted images */ 4445c6dcc57SGao Xiang if (fe->owned_head == Z_EROFS_PCLUSTER_TAIL) 4455c6dcc57SGao Xiang fe->tailpcl = pcl; 446473e15b0SGao Xiang 4475c6dcc57SGao Xiang z_erofs_try_to_claim_pcluster(fe); 4489e579fc1SGao Xiang return 0; 44947e4937aSGao Xiang } 45047e4937aSGao Xiang 45183a386c0SGao Xiang static int z_erofs_register_pcluster(struct z_erofs_decompress_frontend *fe) 45247e4937aSGao Xiang { 45383a386c0SGao Xiang struct erofs_map_blocks *map = &fe->map; 454cecf864dSYue Hu bool ztailpacking = map->m_flags & EROFS_MAP_META; 45547e4937aSGao Xiang struct z_erofs_pcluster *pcl; 45664094a04SGao Xiang struct erofs_workgroup *grp; 45747e4937aSGao Xiang int err; 45847e4937aSGao Xiang 4598f899262SGao Xiang if (!(map->m_flags & EROFS_MAP_ENCODED)) { 4608f899262SGao Xiang DBG_BUGON(1); 4618f899262SGao Xiang return -EFSCORRUPTED; 4628f899262SGao Xiang } 4638f899262SGao Xiang 4649f6cc76eSGao Xiang /* no available pcluster, let's allocate one */ 465cecf864dSYue Hu pcl = z_erofs_alloc_pcluster(ztailpacking ? 1 : 466cecf864dSYue Hu map->m_plen >> PAGE_SHIFT); 4679f6cc76eSGao Xiang if (IS_ERR(pcl)) 4689f6cc76eSGao Xiang return PTR_ERR(pcl); 46947e4937aSGao Xiang 47064094a04SGao Xiang atomic_set(&pcl->obj.refcount, 1); 4718f899262SGao Xiang pcl->algorithmformat = map->m_algorithmformat; 47247e4937aSGao Xiang pcl->length = (map->m_llen << Z_EROFS_PCLUSTER_LENGTH_BIT) | 47347e4937aSGao Xiang (map->m_flags & EROFS_MAP_FULL_MAPPED ? 47447e4937aSGao Xiang Z_EROFS_PCLUSTER_FULL_LENGTH : 0); 47547e4937aSGao Xiang 47647e4937aSGao Xiang /* new pclusters should be claimed as type 1, primary and followed */ 4775c6dcc57SGao Xiang pcl->next = fe->owned_head; 47887ca34a7SGao Xiang pcl->pageofs_out = map->m_la & ~PAGE_MASK; 4795c6dcc57SGao Xiang fe->mode = COLLECT_PRIMARY_FOLLOWED; 48047e4937aSGao Xiang 48147e4937aSGao Xiang /* 48247e4937aSGao Xiang * lock all primary followed works before visible to others 48347e4937aSGao Xiang * and mutex_trylock *never* fails for a new pcluster. 48447e4937aSGao Xiang */ 48587ca34a7SGao Xiang mutex_init(&pcl->lock); 48687ca34a7SGao Xiang DBG_BUGON(!mutex_trylock(&pcl->lock)); 48747e4937aSGao Xiang 488cecf864dSYue Hu if (ztailpacking) { 489cecf864dSYue Hu pcl->obj.index = 0; /* which indicates ztailpacking */ 490cecf864dSYue Hu pcl->pageofs_in = erofs_blkoff(map->m_pa); 491cecf864dSYue Hu pcl->tailpacking_size = map->m_plen; 492cecf864dSYue Hu } else { 493cecf864dSYue Hu pcl->obj.index = map->m_pa >> PAGE_SHIFT; 494cecf864dSYue Hu 49583a386c0SGao Xiang grp = erofs_insert_workgroup(fe->inode->i_sb, &pcl->obj); 49664094a04SGao Xiang if (IS_ERR(grp)) { 49764094a04SGao Xiang err = PTR_ERR(grp); 49864094a04SGao Xiang goto err_out; 49964094a04SGao Xiang } 50064094a04SGao Xiang 50164094a04SGao Xiang if (grp != &pcl->obj) { 5025c6dcc57SGao Xiang fe->pcl = container_of(grp, 503cecf864dSYue Hu struct z_erofs_pcluster, obj); 50464094a04SGao Xiang err = -EEXIST; 50564094a04SGao Xiang goto err_out; 50647e4937aSGao Xiang } 507cecf864dSYue Hu } 50847e4937aSGao Xiang /* used to check tail merging loop due to corrupted images */ 5095c6dcc57SGao Xiang if (fe->owned_head == Z_EROFS_PCLUSTER_TAIL) 5105c6dcc57SGao Xiang fe->tailpcl = pcl; 5115c6dcc57SGao Xiang fe->owned_head = &pcl->next; 5125c6dcc57SGao Xiang fe->pcl = pcl; 5139e579fc1SGao Xiang return 0; 51464094a04SGao Xiang 51564094a04SGao Xiang err_out: 51687ca34a7SGao Xiang mutex_unlock(&pcl->lock); 5179f6cc76eSGao Xiang z_erofs_free_pcluster(pcl); 51864094a04SGao Xiang return err; 51947e4937aSGao Xiang } 52047e4937aSGao Xiang 52183a386c0SGao Xiang static int z_erofs_collector_begin(struct z_erofs_decompress_frontend *fe) 52247e4937aSGao Xiang { 52383a386c0SGao Xiang struct erofs_map_blocks *map = &fe->map; 5240d823b42SGao Xiang struct erofs_workgroup *grp = NULL; 5259e579fc1SGao Xiang int ret; 52647e4937aSGao Xiang 52787ca34a7SGao Xiang DBG_BUGON(fe->pcl); 52847e4937aSGao Xiang 52987ca34a7SGao Xiang /* must be Z_EROFS_PCLUSTER_TAIL or pointed to previous pcluster */ 5305c6dcc57SGao Xiang DBG_BUGON(fe->owned_head == Z_EROFS_PCLUSTER_NIL); 5315c6dcc57SGao Xiang DBG_BUGON(fe->owned_head == Z_EROFS_PCLUSTER_TAIL_CLOSED); 53247e4937aSGao Xiang 5330d823b42SGao Xiang if (!(map->m_flags & EROFS_MAP_META)) { 5340d823b42SGao Xiang grp = erofs_find_workgroup(fe->inode->i_sb, 5350d823b42SGao Xiang map->m_pa >> PAGE_SHIFT); 5360d823b42SGao Xiang } else if ((map->m_pa & ~PAGE_MASK) + map->m_plen > PAGE_SIZE) { 53747e4937aSGao Xiang DBG_BUGON(1); 538cecf864dSYue Hu return -EFSCORRUPTED; 539cecf864dSYue Hu } 54047e4937aSGao Xiang 54164094a04SGao Xiang if (grp) { 5425c6dcc57SGao Xiang fe->pcl = container_of(grp, struct z_erofs_pcluster, obj); 5430d823b42SGao Xiang ret = -EEXIST; 54464094a04SGao Xiang } else { 54583a386c0SGao Xiang ret = z_erofs_register_pcluster(fe); 54664094a04SGao Xiang } 54747e4937aSGao Xiang 5480d823b42SGao Xiang if (ret == -EEXIST) { 54983a386c0SGao Xiang ret = z_erofs_lookup_pcluster(fe); 55064094a04SGao Xiang if (ret) { 5515c6dcc57SGao Xiang erofs_workgroup_put(&fe->pcl->obj); 55264094a04SGao Xiang return ret; 55364094a04SGao Xiang } 5540d823b42SGao Xiang } else if (ret) { 5550d823b42SGao Xiang return ret; 5560d823b42SGao Xiang } 55764094a04SGao Xiang 5585c6dcc57SGao Xiang z_erofs_pagevec_ctor_init(&fe->vector, Z_EROFS_NR_INLINE_PAGEVECS, 55987ca34a7SGao Xiang fe->pcl->pagevec, fe->pcl->vcnt); 56081382f5fSGao Xiang /* since file-backed online pages are traversed in reverse order */ 5615c6dcc57SGao Xiang fe->icpage_ptr = fe->pcl->compressed_pages + 5625c6dcc57SGao Xiang z_erofs_pclusterpages(fe->pcl); 56347e4937aSGao Xiang return 0; 56447e4937aSGao Xiang } 56547e4937aSGao Xiang 56647e4937aSGao Xiang /* 56747e4937aSGao Xiang * keep in mind that no referenced pclusters will be freed 56847e4937aSGao Xiang * only after a RCU grace period. 56947e4937aSGao Xiang */ 57047e4937aSGao Xiang static void z_erofs_rcu_callback(struct rcu_head *head) 57147e4937aSGao Xiang { 57287ca34a7SGao Xiang z_erofs_free_pcluster(container_of(head, 57387ca34a7SGao Xiang struct z_erofs_pcluster, rcu)); 57447e4937aSGao Xiang } 57547e4937aSGao Xiang 57647e4937aSGao Xiang void erofs_workgroup_free_rcu(struct erofs_workgroup *grp) 57747e4937aSGao Xiang { 57847e4937aSGao Xiang struct z_erofs_pcluster *const pcl = 57947e4937aSGao Xiang container_of(grp, struct z_erofs_pcluster, obj); 58047e4937aSGao Xiang 58187ca34a7SGao Xiang call_rcu(&pcl->rcu, z_erofs_rcu_callback); 58247e4937aSGao Xiang } 58347e4937aSGao Xiang 5845c6dcc57SGao Xiang static bool z_erofs_collector_end(struct z_erofs_decompress_frontend *fe) 58547e4937aSGao Xiang { 58687ca34a7SGao Xiang struct z_erofs_pcluster *pcl = fe->pcl; 58747e4937aSGao Xiang 58887ca34a7SGao Xiang if (!pcl) 58947e4937aSGao Xiang return false; 59047e4937aSGao Xiang 5915c6dcc57SGao Xiang z_erofs_pagevec_ctor_exit(&fe->vector, false); 59287ca34a7SGao Xiang mutex_unlock(&pcl->lock); 59347e4937aSGao Xiang 59447e4937aSGao Xiang /* 59547e4937aSGao Xiang * if all pending pages are added, don't hold its reference 59647e4937aSGao Xiang * any longer if the pcluster isn't hosted by ourselves. 59747e4937aSGao Xiang */ 5985c6dcc57SGao Xiang if (fe->mode < COLLECT_PRIMARY_FOLLOWED_NOINPLACE) 59987ca34a7SGao Xiang erofs_workgroup_put(&pcl->obj); 60047e4937aSGao Xiang 60187ca34a7SGao Xiang fe->pcl = NULL; 60247e4937aSGao Xiang return true; 60347e4937aSGao Xiang } 60447e4937aSGao Xiang 60547e4937aSGao Xiang static bool should_alloc_managed_pages(struct z_erofs_decompress_frontend *fe, 60647e4937aSGao Xiang unsigned int cachestrategy, 60747e4937aSGao Xiang erofs_off_t la) 60847e4937aSGao Xiang { 60947e4937aSGao Xiang if (cachestrategy <= EROFS_ZIP_CACHE_DISABLED) 61047e4937aSGao Xiang return false; 61147e4937aSGao Xiang 61247e4937aSGao Xiang if (fe->backmost) 61347e4937aSGao Xiang return true; 61447e4937aSGao Xiang 61547e4937aSGao Xiang return cachestrategy >= EROFS_ZIP_CACHE_READAROUND && 61647e4937aSGao Xiang la < fe->headoffset; 61747e4937aSGao Xiang } 61847e4937aSGao Xiang 61947e4937aSGao Xiang static int z_erofs_do_read_page(struct z_erofs_decompress_frontend *fe, 620eaa9172aSGao Xiang struct page *page, struct page **pagepool) 62147e4937aSGao Xiang { 62247e4937aSGao Xiang struct inode *const inode = fe->inode; 623bda17a45SGao Xiang struct erofs_sb_info *const sbi = EROFS_I_SB(inode); 62447e4937aSGao Xiang struct erofs_map_blocks *const map = &fe->map; 62547e4937aSGao Xiang const loff_t offset = page_offset(page); 626dc76ea8cSGao Xiang bool tight = true; 62747e4937aSGao Xiang 62847e4937aSGao Xiang enum z_erofs_cache_alloctype cache_strategy; 62947e4937aSGao Xiang enum z_erofs_page_type page_type; 63047e4937aSGao Xiang unsigned int cur, end, spiltted, index; 63147e4937aSGao Xiang int err = 0; 63247e4937aSGao Xiang 63347e4937aSGao Xiang /* register locked file pages as online pages in pack */ 63447e4937aSGao Xiang z_erofs_onlinepage_init(page); 63547e4937aSGao Xiang 63647e4937aSGao Xiang spiltted = 0; 63747e4937aSGao Xiang end = PAGE_SIZE; 63847e4937aSGao Xiang repeat: 63947e4937aSGao Xiang cur = end - 1; 64047e4937aSGao Xiang 64139397a46SGao Xiang if (offset + cur < map->m_la || 64239397a46SGao Xiang offset + cur >= map->m_la + map->m_llen) { 64339397a46SGao Xiang erofs_dbg("out-of-range map @ pos %llu", offset + cur); 64447e4937aSGao Xiang 6455c6dcc57SGao Xiang if (z_erofs_collector_end(fe)) 64647e4937aSGao Xiang fe->backmost = false; 64747e4937aSGao Xiang map->m_la = offset + cur; 64847e4937aSGao Xiang map->m_llen = 0; 64947e4937aSGao Xiang err = z_erofs_map_blocks_iter(inode, map, 0); 6508d8a09b0SGao Xiang if (err) 65147e4937aSGao Xiang goto err_out; 65239397a46SGao Xiang } else { 65339397a46SGao Xiang if (fe->pcl) 65439397a46SGao Xiang goto hitted; 65539397a46SGao Xiang /* didn't get a valid pcluster previously (very rare) */ 65639397a46SGao Xiang } 65747e4937aSGao Xiang 6588d8a09b0SGao Xiang if (!(map->m_flags & EROFS_MAP_MAPPED)) 65947e4937aSGao Xiang goto hitted; 66047e4937aSGao Xiang 66183a386c0SGao Xiang err = z_erofs_collector_begin(fe); 6628d8a09b0SGao Xiang if (err) 66347e4937aSGao Xiang goto err_out; 66447e4937aSGao Xiang 6655c6dcc57SGao Xiang if (z_erofs_is_inline_pcluster(fe->pcl)) { 66609c54379SGao Xiang void *mp; 667cecf864dSYue Hu 66809c54379SGao Xiang mp = erofs_read_metabuf(&fe->map.buf, inode->i_sb, 66909c54379SGao Xiang erofs_blknr(map->m_pa), EROFS_NO_KMAP); 67009c54379SGao Xiang if (IS_ERR(mp)) { 67109c54379SGao Xiang err = PTR_ERR(mp); 672cecf864dSYue Hu erofs_err(inode->i_sb, 673cecf864dSYue Hu "failed to get inline page, err %d", err); 674cecf864dSYue Hu goto err_out; 675cecf864dSYue Hu } 67609c54379SGao Xiang get_page(fe->map.buf.page); 6775c6dcc57SGao Xiang WRITE_ONCE(fe->pcl->compressed_pages[0], fe->map.buf.page); 6785c6dcc57SGao Xiang fe->mode = COLLECT_PRIMARY_FOLLOWED_NOINPLACE; 679cecf864dSYue Hu } else { 6806f39d1e1SGao Xiang /* bind cache first when cached decompression is preferred */ 681cecf864dSYue Hu if (should_alloc_managed_pages(fe, sbi->opt.cache_strategy, 682cecf864dSYue Hu map->m_la)) 6831825c8d7SGao Xiang cache_strategy = TRYALLOC; 68447e4937aSGao Xiang else 68547e4937aSGao Xiang cache_strategy = DONTALLOC; 68647e4937aSGao Xiang 6876f39d1e1SGao Xiang z_erofs_bind_cache(fe, cache_strategy, pagepool); 688cecf864dSYue Hu } 68947e4937aSGao Xiang hitted: 690dc76ea8cSGao Xiang /* 691dc76ea8cSGao Xiang * Ensure the current partial page belongs to this submit chain rather 692dc76ea8cSGao Xiang * than other concurrent submit chains or the noio(bypass) chain since 693dc76ea8cSGao Xiang * those chains are handled asynchronously thus the page cannot be used 694dc76ea8cSGao Xiang * for inplace I/O or pagevec (should be processed in strict order.) 695dc76ea8cSGao Xiang */ 6965c6dcc57SGao Xiang tight &= (fe->mode >= COLLECT_PRIMARY_HOOKED && 6975c6dcc57SGao Xiang fe->mode != COLLECT_PRIMARY_FOLLOWED_NOINPLACE); 698dc76ea8cSGao Xiang 69947e4937aSGao Xiang cur = end - min_t(unsigned int, offset + end - map->m_la, end); 7008d8a09b0SGao Xiang if (!(map->m_flags & EROFS_MAP_MAPPED)) { 70147e4937aSGao Xiang zero_user_segment(page, cur, end); 70247e4937aSGao Xiang goto next_part; 70347e4937aSGao Xiang } 70447e4937aSGao Xiang 70547e4937aSGao Xiang /* let's derive page type */ 70647e4937aSGao Xiang page_type = cur ? Z_EROFS_VLE_PAGE_TYPE_HEAD : 70747e4937aSGao Xiang (!spiltted ? Z_EROFS_PAGE_TYPE_EXCLUSIVE : 70847e4937aSGao Xiang (tight ? Z_EROFS_PAGE_TYPE_EXCLUSIVE : 70947e4937aSGao Xiang Z_EROFS_VLE_PAGE_TYPE_TAIL_SHARED)); 71047e4937aSGao Xiang 71147e4937aSGao Xiang if (cur) 7125c6dcc57SGao Xiang tight &= (fe->mode >= COLLECT_PRIMARY_FOLLOWED); 71347e4937aSGao Xiang 71447e4937aSGao Xiang retry: 7155c6dcc57SGao Xiang err = z_erofs_attach_page(fe, page, page_type, 7165c6dcc57SGao Xiang fe->mode >= COLLECT_PRIMARY_FOLLOWED); 7176aaa7b06SGao Xiang /* should allocate an additional short-lived page for pagevec */ 71847e4937aSGao Xiang if (err == -EAGAIN) { 71947e4937aSGao Xiang struct page *const newpage = 720e3f78d5eSChao Yu alloc_page(GFP_NOFS | __GFP_NOFAIL); 72147e4937aSGao Xiang 7226aaa7b06SGao Xiang set_page_private(newpage, Z_EROFS_SHORTLIVED_PAGE); 7235c6dcc57SGao Xiang err = z_erofs_attach_page(fe, newpage, 72486432a6dSGao Xiang Z_EROFS_PAGE_TYPE_EXCLUSIVE, true); 7258d8a09b0SGao Xiang if (!err) 72647e4937aSGao Xiang goto retry; 72747e4937aSGao Xiang } 72847e4937aSGao Xiang 7298d8a09b0SGao Xiang if (err) 73047e4937aSGao Xiang goto err_out; 73147e4937aSGao Xiang 73247e4937aSGao Xiang index = page->index - (map->m_la >> PAGE_SHIFT); 73347e4937aSGao Xiang 73447e4937aSGao Xiang z_erofs_onlinepage_fixup(page, index, true); 73547e4937aSGao Xiang 73647e4937aSGao Xiang /* bump up the number of spiltted parts of a page */ 73747e4937aSGao Xiang ++spiltted; 73847e4937aSGao Xiang /* also update nr_pages */ 73987ca34a7SGao Xiang fe->pcl->nr_pages = max_t(pgoff_t, fe->pcl->nr_pages, index + 1); 74047e4937aSGao Xiang next_part: 74147e4937aSGao Xiang /* can be used for verification */ 74247e4937aSGao Xiang map->m_llen = offset + cur - map->m_la; 74347e4937aSGao Xiang 74447e4937aSGao Xiang end = cur; 74547e4937aSGao Xiang if (end > 0) 74647e4937aSGao Xiang goto repeat; 74747e4937aSGao Xiang 74847e4937aSGao Xiang out: 74947e4937aSGao Xiang z_erofs_onlinepage_endio(page); 75047e4937aSGao Xiang 7514f761fa2SGao Xiang erofs_dbg("%s, finish page: %pK spiltted: %u map->m_llen %llu", 75247e4937aSGao Xiang __func__, page, spiltted, map->m_llen); 75347e4937aSGao Xiang return err; 75447e4937aSGao Xiang 75547e4937aSGao Xiang /* if some error occurred while processing this page */ 75647e4937aSGao Xiang err_out: 75747e4937aSGao Xiang SetPageError(page); 75847e4937aSGao Xiang goto out; 75947e4937aSGao Xiang } 76047e4937aSGao Xiang 76140452ffcSHuang Jianan static bool z_erofs_get_sync_decompress_policy(struct erofs_sb_info *sbi, 76240452ffcSHuang Jianan unsigned int readahead_pages) 76340452ffcSHuang Jianan { 764a2e20a25SMatthew Wilcox (Oracle) /* auto: enable for read_folio, disable for readahead */ 76540452ffcSHuang Jianan if ((sbi->opt.sync_decompress == EROFS_SYNC_DECOMPRESS_AUTO) && 76640452ffcSHuang Jianan !readahead_pages) 76740452ffcSHuang Jianan return true; 76840452ffcSHuang Jianan 76940452ffcSHuang Jianan if ((sbi->opt.sync_decompress == EROFS_SYNC_DECOMPRESS_FORCE_ON) && 77040452ffcSHuang Jianan (readahead_pages <= sbi->opt.max_sync_decompress_pages)) 77140452ffcSHuang Jianan return true; 77240452ffcSHuang Jianan 77340452ffcSHuang Jianan return false; 77440452ffcSHuang Jianan } 77540452ffcSHuang Jianan 7766aaa7b06SGao Xiang static bool z_erofs_page_is_invalidated(struct page *page) 7776aaa7b06SGao Xiang { 7786aaa7b06SGao Xiang return !page->mapping && !z_erofs_is_shortlived_page(page); 7796aaa7b06SGao Xiang } 7806aaa7b06SGao Xiang 781*42fec235SGao Xiang static int z_erofs_parse_out_bvecs(struct z_erofs_pcluster *pcl, 782*42fec235SGao Xiang struct page **pages, struct page **pagepool) 783*42fec235SGao Xiang { 784*42fec235SGao Xiang struct z_erofs_pagevec_ctor ctor; 785*42fec235SGao Xiang enum z_erofs_page_type page_type; 786*42fec235SGao Xiang int i, err = 0; 787*42fec235SGao Xiang 788*42fec235SGao Xiang z_erofs_pagevec_ctor_init(&ctor, Z_EROFS_NR_INLINE_PAGEVECS, 789*42fec235SGao Xiang pcl->pagevec, 0); 790*42fec235SGao Xiang for (i = 0; i < pcl->vcnt; ++i) { 791*42fec235SGao Xiang struct page *page = z_erofs_pagevec_dequeue(&ctor, &page_type); 792*42fec235SGao Xiang unsigned int pagenr; 793*42fec235SGao Xiang 794*42fec235SGao Xiang /* all pages in pagevec ought to be valid */ 795*42fec235SGao Xiang DBG_BUGON(!page); 796*42fec235SGao Xiang DBG_BUGON(z_erofs_page_is_invalidated(page)); 797*42fec235SGao Xiang 798*42fec235SGao Xiang if (z_erofs_put_shortlivedpage(pagepool, page)) 799*42fec235SGao Xiang continue; 800*42fec235SGao Xiang 801*42fec235SGao Xiang if (page_type == Z_EROFS_VLE_PAGE_TYPE_HEAD) 802*42fec235SGao Xiang pagenr = 0; 803*42fec235SGao Xiang else 804*42fec235SGao Xiang pagenr = z_erofs_onlinepage_index(page); 805*42fec235SGao Xiang 806*42fec235SGao Xiang DBG_BUGON(pagenr >= pcl->nr_pages); 807*42fec235SGao Xiang /* 808*42fec235SGao Xiang * currently EROFS doesn't support multiref(dedup), 809*42fec235SGao Xiang * so here erroring out one multiref page. 810*42fec235SGao Xiang */ 811*42fec235SGao Xiang if (pages[pagenr]) { 812*42fec235SGao Xiang DBG_BUGON(1); 813*42fec235SGao Xiang SetPageError(pages[pagenr]); 814*42fec235SGao Xiang z_erofs_onlinepage_endio(pages[pagenr]); 815*42fec235SGao Xiang err = -EFSCORRUPTED; 816*42fec235SGao Xiang } 817*42fec235SGao Xiang pages[pagenr] = page; 818*42fec235SGao Xiang } 819*42fec235SGao Xiang z_erofs_pagevec_ctor_exit(&ctor, true); 820*42fec235SGao Xiang return err; 821*42fec235SGao Xiang } 822*42fec235SGao Xiang 82347e4937aSGao Xiang static int z_erofs_decompress_pcluster(struct super_block *sb, 82447e4937aSGao Xiang struct z_erofs_pcluster *pcl, 825eaa9172aSGao Xiang struct page **pagepool) 82647e4937aSGao Xiang { 82747e4937aSGao Xiang struct erofs_sb_info *const sbi = EROFS_SB(sb); 828cecf864dSYue Hu unsigned int pclusterpages = z_erofs_pclusterpages(pcl); 8299f6cc76eSGao Xiang unsigned int i, inputsize, outputsize, llen, nr_pages; 83047e4937aSGao Xiang struct page *pages_onstack[Z_EROFS_VMAP_ONSTACK_PAGES]; 83147e4937aSGao Xiang struct page **pages, **compressed_pages, *page; 83247e4937aSGao Xiang 83347e4937aSGao Xiang bool overlapped, partial; 83447e4937aSGao Xiang int err; 83547e4937aSGao Xiang 83647e4937aSGao Xiang might_sleep(); 83787ca34a7SGao Xiang DBG_BUGON(!READ_ONCE(pcl->nr_pages)); 83847e4937aSGao Xiang 83987ca34a7SGao Xiang mutex_lock(&pcl->lock); 84087ca34a7SGao Xiang nr_pages = pcl->nr_pages; 84147e4937aSGao Xiang 8428d8a09b0SGao Xiang if (nr_pages <= Z_EROFS_VMAP_ONSTACK_PAGES) { 84347e4937aSGao Xiang pages = pages_onstack; 84447e4937aSGao Xiang } else if (nr_pages <= Z_EROFS_VMAP_GLOBAL_PAGES && 84547e4937aSGao Xiang mutex_trylock(&z_pagemap_global_lock)) { 84647e4937aSGao Xiang pages = z_pagemap_global; 84747e4937aSGao Xiang } else { 84847e4937aSGao Xiang gfp_t gfp_flags = GFP_KERNEL; 84947e4937aSGao Xiang 85047e4937aSGao Xiang if (nr_pages > Z_EROFS_VMAP_GLOBAL_PAGES) 85147e4937aSGao Xiang gfp_flags |= __GFP_NOFAIL; 85247e4937aSGao Xiang 85347e4937aSGao Xiang pages = kvmalloc_array(nr_pages, sizeof(struct page *), 85447e4937aSGao Xiang gfp_flags); 85547e4937aSGao Xiang 85647e4937aSGao Xiang /* fallback to global pagemap for the lowmem scenario */ 8578d8a09b0SGao Xiang if (!pages) { 85847e4937aSGao Xiang mutex_lock(&z_pagemap_global_lock); 85947e4937aSGao Xiang pages = z_pagemap_global; 86047e4937aSGao Xiang } 86147e4937aSGao Xiang } 86247e4937aSGao Xiang 86347e4937aSGao Xiang for (i = 0; i < nr_pages; ++i) 86447e4937aSGao Xiang pages[i] = NULL; 86547e4937aSGao Xiang 866*42fec235SGao Xiang err = z_erofs_parse_out_bvecs(pcl, pages, pagepool); 86747e4937aSGao Xiang 86847e4937aSGao Xiang overlapped = false; 86947e4937aSGao Xiang compressed_pages = pcl->compressed_pages; 87047e4937aSGao Xiang 871cecf864dSYue Hu for (i = 0; i < pclusterpages; ++i) { 87247e4937aSGao Xiang unsigned int pagenr; 87347e4937aSGao Xiang 87447e4937aSGao Xiang page = compressed_pages[i]; 87547e4937aSGao Xiang /* all compressed pages ought to be valid */ 87647e4937aSGao Xiang DBG_BUGON(!page); 87747e4937aSGao Xiang 878cecf864dSYue Hu if (z_erofs_is_inline_pcluster(pcl)) { 879cecf864dSYue Hu if (!PageUptodate(page)) 880cecf864dSYue Hu err = -EIO; 881cecf864dSYue Hu continue; 882cecf864dSYue Hu } 883cecf864dSYue Hu 884cecf864dSYue Hu DBG_BUGON(z_erofs_page_is_invalidated(page)); 8856aaa7b06SGao Xiang if (!z_erofs_is_shortlived_page(page)) { 88647e4937aSGao Xiang if (erofs_page_is_managed(sbi, page)) { 8878d8a09b0SGao Xiang if (!PageUptodate(page)) 88847e4937aSGao Xiang err = -EIO; 88947e4937aSGao Xiang continue; 89047e4937aSGao Xiang } 89147e4937aSGao Xiang 89247e4937aSGao Xiang /* 89347e4937aSGao Xiang * only if non-head page can be selected 89447e4937aSGao Xiang * for inplace decompression 89547e4937aSGao Xiang */ 89647e4937aSGao Xiang pagenr = z_erofs_onlinepage_index(page); 89747e4937aSGao Xiang 89847e4937aSGao Xiang DBG_BUGON(pagenr >= nr_pages); 8998d8a09b0SGao Xiang if (pages[pagenr]) { 90047e4937aSGao Xiang DBG_BUGON(1); 90147e4937aSGao Xiang SetPageError(pages[pagenr]); 90247e4937aSGao Xiang z_erofs_onlinepage_endio(pages[pagenr]); 90347e4937aSGao Xiang err = -EFSCORRUPTED; 90447e4937aSGao Xiang } 90547e4937aSGao Xiang pages[pagenr] = page; 90647e4937aSGao Xiang 90747e4937aSGao Xiang overlapped = true; 90847e4937aSGao Xiang } 90947e4937aSGao Xiang 9106aaa7b06SGao Xiang /* PG_error needs checking for all non-managed pages */ 9118d8a09b0SGao Xiang if (PageError(page)) { 91247e4937aSGao Xiang DBG_BUGON(PageUptodate(page)); 91347e4937aSGao Xiang err = -EIO; 91447e4937aSGao Xiang } 91547e4937aSGao Xiang } 91647e4937aSGao Xiang 9178d8a09b0SGao Xiang if (err) 91847e4937aSGao Xiang goto out; 91947e4937aSGao Xiang 92047e4937aSGao Xiang llen = pcl->length >> Z_EROFS_PCLUSTER_LENGTH_BIT; 92187ca34a7SGao Xiang if (nr_pages << PAGE_SHIFT >= pcl->pageofs_out + llen) { 92247e4937aSGao Xiang outputsize = llen; 92347e4937aSGao Xiang partial = !(pcl->length & Z_EROFS_PCLUSTER_FULL_LENGTH); 92447e4937aSGao Xiang } else { 92587ca34a7SGao Xiang outputsize = (nr_pages << PAGE_SHIFT) - pcl->pageofs_out; 92647e4937aSGao Xiang partial = true; 92747e4937aSGao Xiang } 92847e4937aSGao Xiang 929cecf864dSYue Hu if (z_erofs_is_inline_pcluster(pcl)) 930cecf864dSYue Hu inputsize = pcl->tailpacking_size; 931cecf864dSYue Hu else 932cecf864dSYue Hu inputsize = pclusterpages * PAGE_SIZE; 933cecf864dSYue Hu 93447e4937aSGao Xiang err = z_erofs_decompress(&(struct z_erofs_decompress_req) { 93547e4937aSGao Xiang .sb = sb, 93647e4937aSGao Xiang .in = compressed_pages, 93747e4937aSGao Xiang .out = pages, 938cecf864dSYue Hu .pageofs_in = pcl->pageofs_in, 93987ca34a7SGao Xiang .pageofs_out = pcl->pageofs_out, 9409f6cc76eSGao Xiang .inputsize = inputsize, 94147e4937aSGao Xiang .outputsize = outputsize, 94247e4937aSGao Xiang .alg = pcl->algorithmformat, 94347e4937aSGao Xiang .inplace_io = overlapped, 94447e4937aSGao Xiang .partial_decoding = partial 94547e4937aSGao Xiang }, pagepool); 94647e4937aSGao Xiang 94747e4937aSGao Xiang out: 948cecf864dSYue Hu /* must handle all compressed pages before actual file pages */ 949cecf864dSYue Hu if (z_erofs_is_inline_pcluster(pcl)) { 950cecf864dSYue Hu page = compressed_pages[0]; 951cecf864dSYue Hu WRITE_ONCE(compressed_pages[0], NULL); 952cecf864dSYue Hu put_page(page); 953cecf864dSYue Hu } else { 954cecf864dSYue Hu for (i = 0; i < pclusterpages; ++i) { 95547e4937aSGao Xiang page = compressed_pages[i]; 95647e4937aSGao Xiang 95747e4937aSGao Xiang if (erofs_page_is_managed(sbi, page)) 95847e4937aSGao Xiang continue; 95947e4937aSGao Xiang 9606aaa7b06SGao Xiang /* recycle all individual short-lived pages */ 9616aaa7b06SGao Xiang (void)z_erofs_put_shortlivedpage(pagepool, page); 96247e4937aSGao Xiang WRITE_ONCE(compressed_pages[i], NULL); 96347e4937aSGao Xiang } 964cecf864dSYue Hu } 96547e4937aSGao Xiang 96647e4937aSGao Xiang for (i = 0; i < nr_pages; ++i) { 96747e4937aSGao Xiang page = pages[i]; 96847e4937aSGao Xiang if (!page) 96947e4937aSGao Xiang continue; 97047e4937aSGao Xiang 9716aaa7b06SGao Xiang DBG_BUGON(z_erofs_page_is_invalidated(page)); 97247e4937aSGao Xiang 9736aaa7b06SGao Xiang /* recycle all individual short-lived pages */ 9746aaa7b06SGao Xiang if (z_erofs_put_shortlivedpage(pagepool, page)) 97547e4937aSGao Xiang continue; 97647e4937aSGao Xiang 9778d8a09b0SGao Xiang if (err < 0) 97847e4937aSGao Xiang SetPageError(page); 97947e4937aSGao Xiang 98047e4937aSGao Xiang z_erofs_onlinepage_endio(page); 98147e4937aSGao Xiang } 98247e4937aSGao Xiang 98347e4937aSGao Xiang if (pages == z_pagemap_global) 98447e4937aSGao Xiang mutex_unlock(&z_pagemap_global_lock); 9858d8a09b0SGao Xiang else if (pages != pages_onstack) 98647e4937aSGao Xiang kvfree(pages); 98747e4937aSGao Xiang 98887ca34a7SGao Xiang pcl->nr_pages = 0; 98987ca34a7SGao Xiang pcl->vcnt = 0; 99047e4937aSGao Xiang 99187ca34a7SGao Xiang /* pcluster lock MUST be taken before the following line */ 99247e4937aSGao Xiang WRITE_ONCE(pcl->next, Z_EROFS_PCLUSTER_NIL); 99387ca34a7SGao Xiang mutex_unlock(&pcl->lock); 99447e4937aSGao Xiang return err; 99547e4937aSGao Xiang } 99647e4937aSGao Xiang 9970c638f70SGao Xiang static void z_erofs_decompress_queue(const struct z_erofs_decompressqueue *io, 998eaa9172aSGao Xiang struct page **pagepool) 99947e4937aSGao Xiang { 100047e4937aSGao Xiang z_erofs_next_pcluster_t owned = io->head; 100147e4937aSGao Xiang 100247e4937aSGao Xiang while (owned != Z_EROFS_PCLUSTER_TAIL_CLOSED) { 100347e4937aSGao Xiang struct z_erofs_pcluster *pcl; 100447e4937aSGao Xiang 100547e4937aSGao Xiang /* no possible that 'owned' equals Z_EROFS_WORK_TPTR_TAIL */ 100647e4937aSGao Xiang DBG_BUGON(owned == Z_EROFS_PCLUSTER_TAIL); 100747e4937aSGao Xiang 100847e4937aSGao Xiang /* no possible that 'owned' equals NULL */ 100947e4937aSGao Xiang DBG_BUGON(owned == Z_EROFS_PCLUSTER_NIL); 101047e4937aSGao Xiang 101147e4937aSGao Xiang pcl = container_of(owned, struct z_erofs_pcluster, next); 101247e4937aSGao Xiang owned = READ_ONCE(pcl->next); 101347e4937aSGao Xiang 1014a4b1fab1SGao Xiang z_erofs_decompress_pcluster(io->sb, pcl, pagepool); 101587ca34a7SGao Xiang erofs_workgroup_put(&pcl->obj); 101647e4937aSGao Xiang } 101747e4937aSGao Xiang } 101847e4937aSGao Xiang 10190c638f70SGao Xiang static void z_erofs_decompressqueue_work(struct work_struct *work) 102047e4937aSGao Xiang { 1021a4b1fab1SGao Xiang struct z_erofs_decompressqueue *bgq = 1022a4b1fab1SGao Xiang container_of(work, struct z_erofs_decompressqueue, u.work); 1023eaa9172aSGao Xiang struct page *pagepool = NULL; 102447e4937aSGao Xiang 1025a4b1fab1SGao Xiang DBG_BUGON(bgq->head == Z_EROFS_PCLUSTER_TAIL_CLOSED); 10260c638f70SGao Xiang z_erofs_decompress_queue(bgq, &pagepool); 102747e4937aSGao Xiang 1028eaa9172aSGao Xiang erofs_release_pages(&pagepool); 1029a4b1fab1SGao Xiang kvfree(bgq); 103047e4937aSGao Xiang } 103147e4937aSGao Xiang 10327865827cSGao Xiang static void z_erofs_decompress_kickoff(struct z_erofs_decompressqueue *io, 10337865827cSGao Xiang bool sync, int bios) 10347865827cSGao Xiang { 10357865827cSGao Xiang struct erofs_sb_info *const sbi = EROFS_SB(io->sb); 10367865827cSGao Xiang 10377865827cSGao Xiang /* wake up the caller thread for sync decompression */ 10387865827cSGao Xiang if (sync) { 10397865827cSGao Xiang if (!atomic_add_return(bios, &io->pending_bios)) 104060b30050SHongyu Jin complete(&io->u.done); 104160b30050SHongyu Jin 10427865827cSGao Xiang return; 10437865827cSGao Xiang } 10447865827cSGao Xiang 10457865827cSGao Xiang if (atomic_add_return(bios, &io->pending_bios)) 10467865827cSGao Xiang return; 10477865827cSGao Xiang /* Use workqueue and sync decompression for atomic contexts only */ 10487865827cSGao Xiang if (in_atomic() || irqs_disabled()) { 10497865827cSGao Xiang queue_work(z_erofs_workqueue, &io->u.work); 10507865827cSGao Xiang /* enable sync decompression for readahead */ 10517865827cSGao Xiang if (sbi->opt.sync_decompress == EROFS_SYNC_DECOMPRESS_AUTO) 10527865827cSGao Xiang sbi->opt.sync_decompress = EROFS_SYNC_DECOMPRESS_FORCE_ON; 10537865827cSGao Xiang return; 10547865827cSGao Xiang } 10557865827cSGao Xiang z_erofs_decompressqueue_work(&io->u.work); 10567865827cSGao Xiang } 10577865827cSGao Xiang 105847e4937aSGao Xiang static struct page *pickup_page_for_submission(struct z_erofs_pcluster *pcl, 105947e4937aSGao Xiang unsigned int nr, 1060eaa9172aSGao Xiang struct page **pagepool, 10619f2731d6SGao Xiang struct address_space *mc) 106247e4937aSGao Xiang { 106347e4937aSGao Xiang const pgoff_t index = pcl->obj.index; 10649f2731d6SGao Xiang gfp_t gfp = mapping_gfp_mask(mc); 106547e4937aSGao Xiang bool tocache = false; 106647e4937aSGao Xiang 106747e4937aSGao Xiang struct address_space *mapping; 106847e4937aSGao Xiang struct page *oldpage, *page; 106947e4937aSGao Xiang 107047e4937aSGao Xiang compressed_page_t t; 107147e4937aSGao Xiang int justfound; 107247e4937aSGao Xiang 107347e4937aSGao Xiang repeat: 107447e4937aSGao Xiang page = READ_ONCE(pcl->compressed_pages[nr]); 107547e4937aSGao Xiang oldpage = page; 107647e4937aSGao Xiang 107747e4937aSGao Xiang if (!page) 107847e4937aSGao Xiang goto out_allocpage; 107947e4937aSGao Xiang 108047e4937aSGao Xiang /* process the target tagged pointer */ 108147e4937aSGao Xiang t = tagptr_init(compressed_page_t, page); 108247e4937aSGao Xiang justfound = tagptr_unfold_tags(t); 108347e4937aSGao Xiang page = tagptr_unfold_ptr(t); 108447e4937aSGao Xiang 10851825c8d7SGao Xiang /* 10861825c8d7SGao Xiang * preallocated cached pages, which is used to avoid direct reclaim 10871825c8d7SGao Xiang * otherwise, it will go inplace I/O path instead. 10881825c8d7SGao Xiang */ 10891825c8d7SGao Xiang if (page->private == Z_EROFS_PREALLOCATED_PAGE) { 10901825c8d7SGao Xiang WRITE_ONCE(pcl->compressed_pages[nr], page); 10911825c8d7SGao Xiang set_page_private(page, 0); 10921825c8d7SGao Xiang tocache = true; 10931825c8d7SGao Xiang goto out_tocache; 10941825c8d7SGao Xiang } 109547e4937aSGao Xiang mapping = READ_ONCE(page->mapping); 109647e4937aSGao Xiang 109747e4937aSGao Xiang /* 10986aaa7b06SGao Xiang * file-backed online pages in plcuster are all locked steady, 109947e4937aSGao Xiang * therefore it is impossible for `mapping' to be NULL. 110047e4937aSGao Xiang */ 110147e4937aSGao Xiang if (mapping && mapping != mc) 110247e4937aSGao Xiang /* ought to be unmanaged pages */ 110347e4937aSGao Xiang goto out; 110447e4937aSGao Xiang 11056aaa7b06SGao Xiang /* directly return for shortlived page as well */ 11066aaa7b06SGao Xiang if (z_erofs_is_shortlived_page(page)) 11076aaa7b06SGao Xiang goto out; 11086aaa7b06SGao Xiang 110947e4937aSGao Xiang lock_page(page); 111047e4937aSGao Xiang 111147e4937aSGao Xiang /* only true if page reclaim goes wrong, should never happen */ 111247e4937aSGao Xiang DBG_BUGON(justfound && PagePrivate(page)); 111347e4937aSGao Xiang 111447e4937aSGao Xiang /* the page is still in manage cache */ 111547e4937aSGao Xiang if (page->mapping == mc) { 111647e4937aSGao Xiang WRITE_ONCE(pcl->compressed_pages[nr], page); 111747e4937aSGao Xiang 111847e4937aSGao Xiang ClearPageError(page); 111947e4937aSGao Xiang if (!PagePrivate(page)) { 112047e4937aSGao Xiang /* 112147e4937aSGao Xiang * impossible to be !PagePrivate(page) for 112247e4937aSGao Xiang * the current restriction as well if 112347e4937aSGao Xiang * the page is already in compressed_pages[]. 112447e4937aSGao Xiang */ 112547e4937aSGao Xiang DBG_BUGON(!justfound); 112647e4937aSGao Xiang 112747e4937aSGao Xiang justfound = 0; 112847e4937aSGao Xiang set_page_private(page, (unsigned long)pcl); 112947e4937aSGao Xiang SetPagePrivate(page); 113047e4937aSGao Xiang } 113147e4937aSGao Xiang 113247e4937aSGao Xiang /* no need to submit io if it is already up-to-date */ 113347e4937aSGao Xiang if (PageUptodate(page)) { 113447e4937aSGao Xiang unlock_page(page); 113547e4937aSGao Xiang page = NULL; 113647e4937aSGao Xiang } 113747e4937aSGao Xiang goto out; 113847e4937aSGao Xiang } 113947e4937aSGao Xiang 114047e4937aSGao Xiang /* 114147e4937aSGao Xiang * the managed page has been truncated, it's unsafe to 114247e4937aSGao Xiang * reuse this one, let's allocate a new cache-managed page. 114347e4937aSGao Xiang */ 114447e4937aSGao Xiang DBG_BUGON(page->mapping); 114547e4937aSGao Xiang DBG_BUGON(!justfound); 114647e4937aSGao Xiang 114747e4937aSGao Xiang tocache = true; 114847e4937aSGao Xiang unlock_page(page); 114947e4937aSGao Xiang put_page(page); 115047e4937aSGao Xiang out_allocpage: 11515ddcee1fSGao Xiang page = erofs_allocpage(pagepool, gfp | __GFP_NOFAIL); 11525ddcee1fSGao Xiang if (oldpage != cmpxchg(&pcl->compressed_pages[nr], oldpage, page)) { 1153eaa9172aSGao Xiang erofs_pagepool_add(pagepool, page); 11545ddcee1fSGao Xiang cond_resched(); 11555ddcee1fSGao Xiang goto repeat; 11565ddcee1fSGao Xiang } 11571825c8d7SGao Xiang out_tocache: 1158bf225074SGao Xiang if (!tocache || add_to_page_cache_lru(page, mc, index + nr, gfp)) { 1159bf225074SGao Xiang /* turn into temporary page if fails (1 ref) */ 1160bf225074SGao Xiang set_page_private(page, Z_EROFS_SHORTLIVED_PAGE); 1161bf225074SGao Xiang goto out; 1162a30573b3SGao Xiang } 1163bf225074SGao Xiang attach_page_private(page, pcl); 1164bf225074SGao Xiang /* drop a refcount added by allocpage (then we have 2 refs here) */ 1165bf225074SGao Xiang put_page(page); 1166bf225074SGao Xiang 116747e4937aSGao Xiang out: /* the only exit (for tracing and debugging) */ 116847e4937aSGao Xiang return page; 116947e4937aSGao Xiang } 117047e4937aSGao Xiang 1171a4b1fab1SGao Xiang static struct z_erofs_decompressqueue * 1172a4b1fab1SGao Xiang jobqueue_init(struct super_block *sb, 1173a4b1fab1SGao Xiang struct z_erofs_decompressqueue *fgq, bool *fg) 117447e4937aSGao Xiang { 1175a4b1fab1SGao Xiang struct z_erofs_decompressqueue *q; 117647e4937aSGao Xiang 1177a4b1fab1SGao Xiang if (fg && !*fg) { 1178a4b1fab1SGao Xiang q = kvzalloc(sizeof(*q), GFP_KERNEL | __GFP_NOWARN); 1179a4b1fab1SGao Xiang if (!q) { 1180a4b1fab1SGao Xiang *fg = true; 1181a4b1fab1SGao Xiang goto fg_out; 118247e4937aSGao Xiang } 11830c638f70SGao Xiang INIT_WORK(&q->u.work, z_erofs_decompressqueue_work); 1184a4b1fab1SGao Xiang } else { 1185a4b1fab1SGao Xiang fg_out: 1186a4b1fab1SGao Xiang q = fgq; 118760b30050SHongyu Jin init_completion(&fgq->u.done); 1188a4b1fab1SGao Xiang atomic_set(&fgq->pending_bios, 0); 1189a4b1fab1SGao Xiang } 1190a4b1fab1SGao Xiang q->sb = sb; 1191a4b1fab1SGao Xiang q->head = Z_EROFS_PCLUSTER_TAIL_CLOSED; 1192a4b1fab1SGao Xiang return q; 119347e4937aSGao Xiang } 119447e4937aSGao Xiang 119547e4937aSGao Xiang /* define decompression jobqueue types */ 119647e4937aSGao Xiang enum { 119747e4937aSGao Xiang JQ_BYPASS, 119847e4937aSGao Xiang JQ_SUBMIT, 119947e4937aSGao Xiang NR_JOBQUEUES, 120047e4937aSGao Xiang }; 120147e4937aSGao Xiang 120247e4937aSGao Xiang static void *jobqueueset_init(struct super_block *sb, 1203a4b1fab1SGao Xiang struct z_erofs_decompressqueue *q[], 1204a4b1fab1SGao Xiang struct z_erofs_decompressqueue *fgq, bool *fg) 120547e4937aSGao Xiang { 120647e4937aSGao Xiang /* 120747e4937aSGao Xiang * if managed cache is enabled, bypass jobqueue is needed, 120847e4937aSGao Xiang * no need to read from device for all pclusters in this queue. 120947e4937aSGao Xiang */ 1210a4b1fab1SGao Xiang q[JQ_BYPASS] = jobqueue_init(sb, fgq + JQ_BYPASS, NULL); 1211a4b1fab1SGao Xiang q[JQ_SUBMIT] = jobqueue_init(sb, fgq + JQ_SUBMIT, fg); 121247e4937aSGao Xiang 1213a4b1fab1SGao Xiang return tagptr_cast_ptr(tagptr_fold(tagptr1_t, q[JQ_SUBMIT], *fg)); 121447e4937aSGao Xiang } 121547e4937aSGao Xiang 121647e4937aSGao Xiang static void move_to_bypass_jobqueue(struct z_erofs_pcluster *pcl, 121747e4937aSGao Xiang z_erofs_next_pcluster_t qtail[], 121847e4937aSGao Xiang z_erofs_next_pcluster_t owned_head) 121947e4937aSGao Xiang { 122047e4937aSGao Xiang z_erofs_next_pcluster_t *const submit_qtail = qtail[JQ_SUBMIT]; 122147e4937aSGao Xiang z_erofs_next_pcluster_t *const bypass_qtail = qtail[JQ_BYPASS]; 122247e4937aSGao Xiang 122347e4937aSGao Xiang DBG_BUGON(owned_head == Z_EROFS_PCLUSTER_TAIL_CLOSED); 122447e4937aSGao Xiang if (owned_head == Z_EROFS_PCLUSTER_TAIL) 122547e4937aSGao Xiang owned_head = Z_EROFS_PCLUSTER_TAIL_CLOSED; 122647e4937aSGao Xiang 122747e4937aSGao Xiang WRITE_ONCE(pcl->next, Z_EROFS_PCLUSTER_TAIL_CLOSED); 122847e4937aSGao Xiang 122947e4937aSGao Xiang WRITE_ONCE(*submit_qtail, owned_head); 123047e4937aSGao Xiang WRITE_ONCE(*bypass_qtail, &pcl->next); 123147e4937aSGao Xiang 123247e4937aSGao Xiang qtail[JQ_BYPASS] = &pcl->next; 123347e4937aSGao Xiang } 123447e4937aSGao Xiang 12357865827cSGao Xiang static void z_erofs_decompressqueue_endio(struct bio *bio) 12367865827cSGao Xiang { 12377865827cSGao Xiang tagptr1_t t = tagptr_init(tagptr1_t, bio->bi_private); 12387865827cSGao Xiang struct z_erofs_decompressqueue *q = tagptr_unfold_ptr(t); 12397865827cSGao Xiang blk_status_t err = bio->bi_status; 12407865827cSGao Xiang struct bio_vec *bvec; 12417865827cSGao Xiang struct bvec_iter_all iter_all; 12427865827cSGao Xiang 12437865827cSGao Xiang bio_for_each_segment_all(bvec, bio, iter_all) { 12447865827cSGao Xiang struct page *page = bvec->bv_page; 12457865827cSGao Xiang 12467865827cSGao Xiang DBG_BUGON(PageUptodate(page)); 12477865827cSGao Xiang DBG_BUGON(z_erofs_page_is_invalidated(page)); 12487865827cSGao Xiang 12497865827cSGao Xiang if (err) 12507865827cSGao Xiang SetPageError(page); 12517865827cSGao Xiang 12527865827cSGao Xiang if (erofs_page_is_managed(EROFS_SB(q->sb), page)) { 12537865827cSGao Xiang if (!err) 12547865827cSGao Xiang SetPageUptodate(page); 12557865827cSGao Xiang unlock_page(page); 12567865827cSGao Xiang } 12577865827cSGao Xiang } 12587865827cSGao Xiang z_erofs_decompress_kickoff(q, tagptr_unfold_tags(t), -1); 12597865827cSGao Xiang bio_put(bio); 12607865827cSGao Xiang } 12617865827cSGao Xiang 126283a386c0SGao Xiang static void z_erofs_submit_queue(struct z_erofs_decompress_frontend *f, 1263eaa9172aSGao Xiang struct page **pagepool, 1264a4b1fab1SGao Xiang struct z_erofs_decompressqueue *fgq, 1265a4b1fab1SGao Xiang bool *force_fg) 126647e4937aSGao Xiang { 126783a386c0SGao Xiang struct super_block *sb = f->inode->i_sb; 126883a386c0SGao Xiang struct address_space *mc = MNGD_MAPPING(EROFS_SB(sb)); 126947e4937aSGao Xiang z_erofs_next_pcluster_t qtail[NR_JOBQUEUES]; 1270a4b1fab1SGao Xiang struct z_erofs_decompressqueue *q[NR_JOBQUEUES]; 127147e4937aSGao Xiang void *bi_private; 12725c6dcc57SGao Xiang z_erofs_next_pcluster_t owned_head = f->owned_head; 1273dfeab2e9SGao Xiang /* bio is NULL initially, so no need to initialize last_{index,bdev} */ 12743f649ab7SKees Cook pgoff_t last_index; 1275dfeab2e9SGao Xiang struct block_device *last_bdev; 12761e4a2955SGao Xiang unsigned int nr_bios = 0; 12771e4a2955SGao Xiang struct bio *bio = NULL; 127847e4937aSGao Xiang 1279a4b1fab1SGao Xiang bi_private = jobqueueset_init(sb, q, fgq, force_fg); 1280a4b1fab1SGao Xiang qtail[JQ_BYPASS] = &q[JQ_BYPASS]->head; 1281a4b1fab1SGao Xiang qtail[JQ_SUBMIT] = &q[JQ_SUBMIT]->head; 128247e4937aSGao Xiang 128347e4937aSGao Xiang /* by default, all need io submission */ 128447e4937aSGao Xiang q[JQ_SUBMIT]->head = owned_head; 128547e4937aSGao Xiang 128647e4937aSGao Xiang do { 1287dfeab2e9SGao Xiang struct erofs_map_dev mdev; 128847e4937aSGao Xiang struct z_erofs_pcluster *pcl; 12891e4a2955SGao Xiang pgoff_t cur, end; 12901e4a2955SGao Xiang unsigned int i = 0; 12911e4a2955SGao Xiang bool bypass = true; 129247e4937aSGao Xiang 129347e4937aSGao Xiang /* no possible 'owned_head' equals the following */ 129447e4937aSGao Xiang DBG_BUGON(owned_head == Z_EROFS_PCLUSTER_TAIL_CLOSED); 129547e4937aSGao Xiang DBG_BUGON(owned_head == Z_EROFS_PCLUSTER_NIL); 129647e4937aSGao Xiang 129747e4937aSGao Xiang pcl = container_of(owned_head, struct z_erofs_pcluster, next); 129847e4937aSGao Xiang 1299cecf864dSYue Hu /* close the main owned chain at first */ 1300cecf864dSYue Hu owned_head = cmpxchg(&pcl->next, Z_EROFS_PCLUSTER_TAIL, 1301cecf864dSYue Hu Z_EROFS_PCLUSTER_TAIL_CLOSED); 1302cecf864dSYue Hu if (z_erofs_is_inline_pcluster(pcl)) { 1303cecf864dSYue Hu move_to_bypass_jobqueue(pcl, qtail, owned_head); 1304cecf864dSYue Hu continue; 1305cecf864dSYue Hu } 1306cecf864dSYue Hu 1307dfeab2e9SGao Xiang /* no device id here, thus it will always succeed */ 1308dfeab2e9SGao Xiang mdev = (struct erofs_map_dev) { 1309dfeab2e9SGao Xiang .m_pa = blknr_to_addr(pcl->obj.index), 1310dfeab2e9SGao Xiang }; 1311dfeab2e9SGao Xiang (void)erofs_map_dev(sb, &mdev); 1312dfeab2e9SGao Xiang 1313dfeab2e9SGao Xiang cur = erofs_blknr(mdev.m_pa); 13149f6cc76eSGao Xiang end = cur + pcl->pclusterpages; 131547e4937aSGao Xiang 13161e4a2955SGao Xiang do { 13171e4a2955SGao Xiang struct page *page; 131847e4937aSGao Xiang 13191e4a2955SGao Xiang page = pickup_page_for_submission(pcl, i++, pagepool, 132083a386c0SGao Xiang mc); 13211e4a2955SGao Xiang if (!page) 13221e4a2955SGao Xiang continue; 132347e4937aSGao Xiang 1324dfeab2e9SGao Xiang if (bio && (cur != last_index + 1 || 1325dfeab2e9SGao Xiang last_bdev != mdev.m_bdev)) { 132647e4937aSGao Xiang submit_bio_retry: 132794e4e153SGao Xiang submit_bio(bio); 132847e4937aSGao Xiang bio = NULL; 132947e4937aSGao Xiang } 133047e4937aSGao Xiang 133147e4937aSGao Xiang if (!bio) { 133207888c66SChristoph Hellwig bio = bio_alloc(mdev.m_bdev, BIO_MAX_VECS, 133307888c66SChristoph Hellwig REQ_OP_READ, GFP_NOIO); 13340c638f70SGao Xiang bio->bi_end_io = z_erofs_decompressqueue_endio; 1335dfeab2e9SGao Xiang 1336dfeab2e9SGao Xiang last_bdev = mdev.m_bdev; 13371e4a2955SGao Xiang bio->bi_iter.bi_sector = (sector_t)cur << 1338a5c0b780SGao Xiang LOG_SECTORS_PER_BLOCK; 1339a5c0b780SGao Xiang bio->bi_private = bi_private; 13406ea5aad3SGao Xiang if (f->readahead) 13416ea5aad3SGao Xiang bio->bi_opf |= REQ_RAHEAD; 134247e4937aSGao Xiang ++nr_bios; 134347e4937aSGao Xiang } 134447e4937aSGao Xiang 13456c3e485eSGao Xiang if (bio_add_page(bio, page, PAGE_SIZE, 0) < PAGE_SIZE) 134647e4937aSGao Xiang goto submit_bio_retry; 134747e4937aSGao Xiang 13481e4a2955SGao Xiang last_index = cur; 13491e4a2955SGao Xiang bypass = false; 13501e4a2955SGao Xiang } while (++cur < end); 135147e4937aSGao Xiang 13521e4a2955SGao Xiang if (!bypass) 135347e4937aSGao Xiang qtail[JQ_SUBMIT] = &pcl->next; 135447e4937aSGao Xiang else 135547e4937aSGao Xiang move_to_bypass_jobqueue(pcl, qtail, owned_head); 135647e4937aSGao Xiang } while (owned_head != Z_EROFS_PCLUSTER_TAIL); 135747e4937aSGao Xiang 135847e4937aSGao Xiang if (bio) 135994e4e153SGao Xiang submit_bio(bio); 136047e4937aSGao Xiang 1361587a67b7SGao Xiang /* 1362587a67b7SGao Xiang * although background is preferred, no one is pending for submission. 1363587a67b7SGao Xiang * don't issue workqueue for decompression but drop it directly instead. 1364587a67b7SGao Xiang */ 1365587a67b7SGao Xiang if (!*force_fg && !nr_bios) { 1366587a67b7SGao Xiang kvfree(q[JQ_SUBMIT]); 13671e4a2955SGao Xiang return; 1368587a67b7SGao Xiang } 1369a4b1fab1SGao Xiang z_erofs_decompress_kickoff(q[JQ_SUBMIT], *force_fg, nr_bios); 137047e4937aSGao Xiang } 137147e4937aSGao Xiang 137283a386c0SGao Xiang static void z_erofs_runqueue(struct z_erofs_decompress_frontend *f, 1373eaa9172aSGao Xiang struct page **pagepool, bool force_fg) 137447e4937aSGao Xiang { 1375a4b1fab1SGao Xiang struct z_erofs_decompressqueue io[NR_JOBQUEUES]; 137647e4937aSGao Xiang 13775c6dcc57SGao Xiang if (f->owned_head == Z_EROFS_PCLUSTER_TAIL) 137847e4937aSGao Xiang return; 137983a386c0SGao Xiang z_erofs_submit_queue(f, pagepool, io, &force_fg); 138047e4937aSGao Xiang 13810c638f70SGao Xiang /* handle bypass queue (no i/o pclusters) immediately */ 13820c638f70SGao Xiang z_erofs_decompress_queue(&io[JQ_BYPASS], pagepool); 138347e4937aSGao Xiang 138447e4937aSGao Xiang if (!force_fg) 138547e4937aSGao Xiang return; 138647e4937aSGao Xiang 138747e4937aSGao Xiang /* wait until all bios are completed */ 138860b30050SHongyu Jin wait_for_completion_io(&io[JQ_SUBMIT].u.done); 138947e4937aSGao Xiang 13900c638f70SGao Xiang /* handle synchronous decompress queue in the caller context */ 13910c638f70SGao Xiang z_erofs_decompress_queue(&io[JQ_SUBMIT], pagepool); 139247e4937aSGao Xiang } 139347e4937aSGao Xiang 139438629291SGao Xiang /* 139538629291SGao Xiang * Since partial uptodate is still unimplemented for now, we have to use 139638629291SGao Xiang * approximate readmore strategies as a start. 139738629291SGao Xiang */ 139838629291SGao Xiang static void z_erofs_pcluster_readmore(struct z_erofs_decompress_frontend *f, 139938629291SGao Xiang struct readahead_control *rac, 140038629291SGao Xiang erofs_off_t end, 1401eaa9172aSGao Xiang struct page **pagepool, 140238629291SGao Xiang bool backmost) 140338629291SGao Xiang { 140438629291SGao Xiang struct inode *inode = f->inode; 140538629291SGao Xiang struct erofs_map_blocks *map = &f->map; 140638629291SGao Xiang erofs_off_t cur; 140738629291SGao Xiang int err; 140838629291SGao Xiang 140938629291SGao Xiang if (backmost) { 141038629291SGao Xiang map->m_la = end; 1411622ceaddSGao Xiang err = z_erofs_map_blocks_iter(inode, map, 1412622ceaddSGao Xiang EROFS_GET_BLOCKS_READMORE); 141338629291SGao Xiang if (err) 141438629291SGao Xiang return; 141538629291SGao Xiang 141638629291SGao Xiang /* expend ra for the trailing edge if readahead */ 141738629291SGao Xiang if (rac) { 141838629291SGao Xiang loff_t newstart = readahead_pos(rac); 141938629291SGao Xiang 142038629291SGao Xiang cur = round_up(map->m_la + map->m_llen, PAGE_SIZE); 142138629291SGao Xiang readahead_expand(rac, newstart, cur - newstart); 142238629291SGao Xiang return; 142338629291SGao Xiang } 142438629291SGao Xiang end = round_up(end, PAGE_SIZE); 142538629291SGao Xiang } else { 142638629291SGao Xiang end = round_up(map->m_la, PAGE_SIZE); 142738629291SGao Xiang 142838629291SGao Xiang if (!map->m_llen) 142938629291SGao Xiang return; 143038629291SGao Xiang } 143138629291SGao Xiang 143238629291SGao Xiang cur = map->m_la + map->m_llen - 1; 143338629291SGao Xiang while (cur >= end) { 143438629291SGao Xiang pgoff_t index = cur >> PAGE_SHIFT; 143538629291SGao Xiang struct page *page; 143638629291SGao Xiang 143738629291SGao Xiang page = erofs_grab_cache_page_nowait(inode->i_mapping, index); 1438aa793b46SGao Xiang if (page) { 143938629291SGao Xiang if (PageUptodate(page)) { 144038629291SGao Xiang unlock_page(page); 1441aa793b46SGao Xiang } else { 144238629291SGao Xiang err = z_erofs_do_read_page(f, page, pagepool); 144338629291SGao Xiang if (err) 144438629291SGao Xiang erofs_err(inode->i_sb, 144538629291SGao Xiang "readmore error at page %lu @ nid %llu", 144638629291SGao Xiang index, EROFS_I(inode)->nid); 1447aa793b46SGao Xiang } 144838629291SGao Xiang put_page(page); 1449aa793b46SGao Xiang } 1450aa793b46SGao Xiang 145138629291SGao Xiang if (cur < PAGE_SIZE) 145238629291SGao Xiang break; 145338629291SGao Xiang cur = (index << PAGE_SHIFT) - 1; 145438629291SGao Xiang } 145538629291SGao Xiang } 145638629291SGao Xiang 1457a2e20a25SMatthew Wilcox (Oracle) static int z_erofs_read_folio(struct file *file, struct folio *folio) 145847e4937aSGao Xiang { 1459a2e20a25SMatthew Wilcox (Oracle) struct page *page = &folio->page; 146047e4937aSGao Xiang struct inode *const inode = page->mapping->host; 146140452ffcSHuang Jianan struct erofs_sb_info *const sbi = EROFS_I_SB(inode); 146247e4937aSGao Xiang struct z_erofs_decompress_frontend f = DECOMPRESS_FRONTEND_INIT(inode); 1463eaa9172aSGao Xiang struct page *pagepool = NULL; 146447e4937aSGao Xiang int err; 146547e4937aSGao Xiang 146647e4937aSGao Xiang trace_erofs_readpage(page, false); 146747e4937aSGao Xiang f.headoffset = (erofs_off_t)page->index << PAGE_SHIFT; 146847e4937aSGao Xiang 146938629291SGao Xiang z_erofs_pcluster_readmore(&f, NULL, f.headoffset + PAGE_SIZE - 1, 147038629291SGao Xiang &pagepool, true); 14711825c8d7SGao Xiang err = z_erofs_do_read_page(&f, page, &pagepool); 147238629291SGao Xiang z_erofs_pcluster_readmore(&f, NULL, 0, &pagepool, false); 147338629291SGao Xiang 14745c6dcc57SGao Xiang (void)z_erofs_collector_end(&f); 147547e4937aSGao Xiang 147647e4937aSGao Xiang /* if some compressed cluster ready, need submit them anyway */ 147783a386c0SGao Xiang z_erofs_runqueue(&f, &pagepool, 147840452ffcSHuang Jianan z_erofs_get_sync_decompress_policy(sbi, 0)); 147947e4937aSGao Xiang 148047e4937aSGao Xiang if (err) 14814f761fa2SGao Xiang erofs_err(inode->i_sb, "failed to read, err [%d]", err); 148247e4937aSGao Xiang 148309c54379SGao Xiang erofs_put_metabuf(&f.map.buf); 1484eaa9172aSGao Xiang erofs_release_pages(&pagepool); 148547e4937aSGao Xiang return err; 148647e4937aSGao Xiang } 148747e4937aSGao Xiang 14880615090cSMatthew Wilcox (Oracle) static void z_erofs_readahead(struct readahead_control *rac) 148947e4937aSGao Xiang { 14900615090cSMatthew Wilcox (Oracle) struct inode *const inode = rac->mapping->host; 149147e4937aSGao Xiang struct erofs_sb_info *const sbi = EROFS_I_SB(inode); 149247e4937aSGao Xiang struct z_erofs_decompress_frontend f = DECOMPRESS_FRONTEND_INIT(inode); 1493eaa9172aSGao Xiang struct page *pagepool = NULL, *head = NULL, *page; 149438629291SGao Xiang unsigned int nr_pages; 149547e4937aSGao Xiang 14966ea5aad3SGao Xiang f.readahead = true; 14970615090cSMatthew Wilcox (Oracle) f.headoffset = readahead_pos(rac); 149847e4937aSGao Xiang 149938629291SGao Xiang z_erofs_pcluster_readmore(&f, rac, f.headoffset + 150038629291SGao Xiang readahead_length(rac) - 1, &pagepool, true); 150138629291SGao Xiang nr_pages = readahead_count(rac); 150238629291SGao Xiang trace_erofs_readpages(inode, readahead_index(rac), nr_pages, false); 150338629291SGao Xiang 15040615090cSMatthew Wilcox (Oracle) while ((page = readahead_page(rac))) { 150547e4937aSGao Xiang set_page_private(page, (unsigned long)head); 150647e4937aSGao Xiang head = page; 150747e4937aSGao Xiang } 150847e4937aSGao Xiang 150947e4937aSGao Xiang while (head) { 151047e4937aSGao Xiang struct page *page = head; 151147e4937aSGao Xiang int err; 151247e4937aSGao Xiang 151347e4937aSGao Xiang /* traversal in reverse order */ 151447e4937aSGao Xiang head = (void *)page_private(page); 151547e4937aSGao Xiang 15161825c8d7SGao Xiang err = z_erofs_do_read_page(&f, page, &pagepool); 1517a5876e24SGao Xiang if (err) 15184f761fa2SGao Xiang erofs_err(inode->i_sb, 15194f761fa2SGao Xiang "readahead error at page %lu @ nid %llu", 15204f761fa2SGao Xiang page->index, EROFS_I(inode)->nid); 152147e4937aSGao Xiang put_page(page); 152247e4937aSGao Xiang } 152338629291SGao Xiang z_erofs_pcluster_readmore(&f, rac, 0, &pagepool, false); 15245c6dcc57SGao Xiang (void)z_erofs_collector_end(&f); 152547e4937aSGao Xiang 152683a386c0SGao Xiang z_erofs_runqueue(&f, &pagepool, 152740452ffcSHuang Jianan z_erofs_get_sync_decompress_policy(sbi, nr_pages)); 152809c54379SGao Xiang erofs_put_metabuf(&f.map.buf); 1529eaa9172aSGao Xiang erofs_release_pages(&pagepool); 153047e4937aSGao Xiang } 153147e4937aSGao Xiang 15320c638f70SGao Xiang const struct address_space_operations z_erofs_aops = { 1533a2e20a25SMatthew Wilcox (Oracle) .read_folio = z_erofs_read_folio, 15340615090cSMatthew Wilcox (Oracle) .readahead = z_erofs_readahead, 153547e4937aSGao Xiang }; 1536